Machine vision techniques, such as simultaneous localization and mapping (SLAM), augmented reality (AR), and virtual reality (VR), often rely on the identification of objects within the local environment of a device through the analysis of imagery of the local environment captured by the device. This analysis conventionally is performed by a graphics processing unit (GPU) on a frame-by-frame basis, which can introduce a significant lag between image capture and feature identification, which in turn can severely impact the perceived responsiveness of the device to changes in the position or orientation of the device or to changes in the local environment.